Bump GPUCompiler.

[maleadt]

Should make 1.9 CI work again.

[maleadt]

@vchuravy You already tested -rc1; did you encounter these regressions?
beta4 worked fine: https://buildkite.com/julialang/cuda-dot-jl/builds/3631#0186b5ba-3223-48f3-b6f4-6f6a40416eaa

cc @KristofferC, just so you know that something may be up with -rc1/GPU back-ends/kwcall.

[vchuravy]

Sorry I had only ran the GPUCompiler tests

[maleadt]

Smells like JuliaLang/julia#48097

[maleadt]

09757f7c83f91b8428f8229375f77810fa829683 is the first bad commit
commit 09757f7c83f91b8428f8229375f77810fa829683
Author: Shuhei Kadowaki <40514306+aviatesk@users.noreply.github.com>
Date:   Wed Mar 1 19:07:59 2023 +0900

    effects: taint `nonoverlayed` when bailing out inference (#48838)

@aviatesk Any thoughts on which JuliaLang/julia#48838 broke kwcall in GPU code again? This due to the presence of overlay methods (seeing how we overlay Core.throw_inexacterror), but we can't fix the underlying issue in time for 1.9, so that's why I did JuliaLang/julia#48116 when last looking at this.

Also, what's the reason you backported that PR? It doesn't reference anything it fixes, nor ever had the backport label (or a discussion on whether it should). So maybe the easiest "fix" is to revert that change?

x-ref JuliaGPU/GPUCompiler.jl#384

[maleadt]

Merging this and moving the discussion to an issue.